The AI Update will be a little different this week, and in the weeks to come. If you have been following this column for a while – first of all, thank you – you know that the ID Tech team, and myself in particular, have been trying to keep track of some of the major developments at the intersection of identity security and AI advances, as well as tracking some of the most important developments unfolding in AI more broadly.
Much of that effort is also now being undertaken at our newly launched sister site/newsletter, Control Plane, which has been focused not only on identity security but also on AI risk more broadly, as well as geopolitical tensions and the broader infrastructure buildout. All of these are related, and it looks like there will be plenty for us to cover at Control Plane for quite some time – so please follow along with us there, if those topics interest you.
Here, in this pivotal edition of the AI Update, I am going to try to offer a snapshot of the big picture issues in AI and identity security as things stand now, on the eve of America’s 250th birthday.
It’s a poignant historical marker. The United States is, by all accounts, the current leader in cutting-edge AI. It has the frontier labs (Elon Musk prefers to simply call them “companies”, but I think they’re more than that), and they have the most sophisticated models, after years of intense competition with Chinese labs, which have shifted to a focus on open source tech. And for that reason, the US government is starting to tighten its grip on its domestic firms. (So has China, but in a slightly different way and for different reasons.)
At Control Plane, a running theme in my newsletters has been America’s gradual but seemingly inevitable trajectory toward nationalization of the leading AI labs. That’s a prediction that was made by former OpenAI researcher Leopold Aschenbrenner back in early 2024, in a long discussion with podcaster Dwarkesh Patel. And his prediction has turned out to be remarkably close, in its details, to what has unfolded since then. Essentially, he argued that AI represents a technological advancement that is so profound that it can be compared to that of nuclear weapons, and that it similarly poses catastrophic risks. As this becomes increasingly clear to observers in the US government, he argued, they will also realize that it must be controlled under government oversight.
Fast forward to early 2026. Anthropic announced a new AI model that it claimed was so dangerous, from a cybersecurity perspective, that it would not release it widely until it had been stress-tested by key partners in various sectors. Then, when it released a modified version of the model for its paying customers, the US government essentially ordered it to be taken offline by imposing strict export controls.
As of this week, access to that model – Fable 5 – has now been restored, after certain additional safeguards were implemented and after Anthropic pleaded to work with the government in a more collaborative relationship going forward. OpenAI, its arch-rival, has meanwhile agreed to stagger the rollout of its newest model to satisfy government concerns.
This isn’t quite nationalization, but it’s obvious that the government is seeking to exert greater control over these companies and their technology. And the models are only getting smarter and more capable. Yes, the US government has ultimately allowed Fable 5 to be released. But what about Fable 6? Fable 7? At what point does it simply become too dangerous to release – and too powerful to leave in the hands of a private company?
These concerns are about cybersecurity in general, and we can infer the identity security implications specifically. We are already seeing how AI is being used to launch identity attacks at scale, helping to automate processes that let fraudsters send massive numbers of phishing emails to potential victims, for example. And even if the US government soon exerts full control over America’s frontier labs, this cybersecurity issue will persist, not least because of China and its labs’ open source models. Technically, usually, they are “open weight”, to be clear – but the general idea is that anyone can see how they work and can modify them to their own purposes, like open source software.
That’s going to be a problem as the open weight models continue to get better. Researchers and analysts generally believe that the intelligence and capabilities of open weight AI models lag behind the cutting-edge frontier tech by a handful of months – maybe six to nine. Which means that around the end of this year, we might expect to see Fable-level intelligence from open weight models. And that means fraudsters and other criminals will have access to superintelligent hacking machines that can be customized to serve their purposes.
That is the worrying reality of the current trajectory. For now, though, we can salute America on its 250th birthday for once again producing technological marvels, and for starting – in a fumbling, ad hoc way – to try to figure out how to control them.
–
By Alex Perala
